03 · How I got here
One layer deeper
Computer science, then data engineering, then analytics, and now the systems that make machine learning fast. Each step started with a question the last one couldn’t answer.
Apr 2026 – now
OngoingSelf-directed
Writing GPU kernels
FlashAttention made me wonder why the same maths can run at wildly different speeds. So I started writing Triton kernels on a laptop GPU and profiling them to find out.
- fused bias + GELU + dropout vs. PyTorch
- 3.3×
- matmul once tensor cores kicked in (TF32)
- 1.45×
- occupancy… and 73% slower. A useful negative result.
- 97%
Mar 2026
Self-taught
Teaching myself CUDA
Before writing fast kernels I wanted to understand the machine they run on: threads, blocks, warps and the memory hierarchy. I learned it from NVIDIA's own documentation and YouTube.
- CUDA fundamentals
- Triton kernels
- Profiling in Nsight
2025
Side project · pit stop
The F1 project
I started it with no plan beyond “sharpen my skills on something new”, scraped 75 seasons of Formula 1 into Postgres, and modelled it in Power BI. Two questions I keep asking it: how much does where you start decide where you finish, and who has owned each era?
Where the winner started
Percentage of races won from each starting position, by era Era P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11+ Next gen 2026– A new era of sustainable fuels began this year. Not in my data yet. Ground effect 2022–25 55 21 11 4 1 2 1 · 1 1 2 Turbo-hybrid 2014–21 53 25 11 4 2 3 · 1 · 2 1 V8 2006–13 51 27 9 5 3 1 1 1 · · 2 % of each era’s races won from each starting slot. Pole has edged up every era, from 51% to 53% to 55%. The furthest-back winner since 2006 started 17th (Verstappen, Brazil 2024). Who owned each era
Cumulative Grand Prix wins, 1950–2025, with the three eras shaded. Mercedes took 111 of 160 hybrid-era races; Red Bull won 55 of 92 in the ground-effect era.